TL;DR
This paper proposes that language change, specifically in noun classifier systems, is primarily driven by child language acquisition rather than functional pressures like ambiguity avoidance. Using computational modeling and corpus analysis of Mandarin and Cantonese, they demonstrate that a neutral, acquisition-driven model can explain the observed generalization trends in classifier systems over time.
What Problem Does This Paper Address?
Classifier systems, common in many languages (especially in Asia), are complex grammatical tools that categorize nouns based on semantic features (e.g., shape, animacy). These systems are known for their variability, ambiguity, and constant evolution. While previous research has explored the historical development and acquisition of classifiers, there's been debate about the drivers of this change. Specifically, do classifiers evolve to better disambiguate meaning (a functional explanation), or do they change due to inherent properties of language learning and transmission (a neutral explanation)? This paper aims to disentangle these possibilities by modeling classifier change as a population-level outcome of child language acquisition.
Methods
The authors employ a two-pronged approach:
- Corpus Analysis: They conducted a quantitative analysis of child-directed speech (CDS) in Mandarin and Cantonese. The goal was to see if Mandarin, which is expected to have more homophony (words sounding alike), shows more evidence of classifiers being used to avoid ambiguity compared to Cantonese.
- Computational Simulation: They developed a population-level transmission model based on the "Tolerance Principle" (TP), a model of productivity learning that has been successful in explaining child language acquisition patterns. This model simulates how a population of "learners" (children) acquire and transmit a classifier system from "adults" over many iterations. The simulation explores whether acquisition alone, without explicit functional pressures, can drive classifier generalization.
Datasets
- Mandarin and Cantonese Child-Directed Speech Corpora: The authors used data from the CHILDES database (specifically, the R conversion by Sanchez et al., 2019) for Mandarin and Cantonese child-directed speech. They excluded one corpus (Erbaugh) that could not be retrieved.
- No new benchmarks were introduced.
Key Insights and Ideas
- Acquisition as the Primary Driver: The central idea is that language change, particularly in classifier systems, is a natural consequence of how children learn language, rather than adults actively trying to fix functional problems like ambiguity.
- Neutral Change Hypothesis Supported: The study provides evidence that a "neutral" model of change, driven by learning and input sparsity, is sufficient to explain broad trends in classifier generalization.
- Ambiguity Avoidance Less Likely: The corpus analysis suggests that Mandarin classifiers don't significantly disambiguate homophones more than Cantonese classifiers, despite Mandarin having more homophonous words. This weakens the argument that functional pressures like ambiguity avoidance are the main drivers of classifier change.
- Generalization is Key: The simulation shows that classifiers tend to generalize over time, losing their specific semantic features, which aligns with historical observations of classifier systems.
- The Tolerance Principle (TP) is Effective: The TP, a model of learning that accounts for U-shaped learning trajectories (memorization -> over-generalization), successfully models the observed developmental and diachronic patterns in classifier systems.
Key Results
- Corpus Analysis: While Mandarin CDS showed more homophonous word types than Cantonese, its classifiers disambiguated homophones significantly less often. This contrasts with polysyllabicity in Mandarin, which does show a trend consistent with homophony avoidance.
- Simulation Results:
- The population-level simulation consistently showed a general trend of classifier generalization over time, with the average number of features per classifier decreasing.
- This trend held across a wide range of parameter settings, suggesting it's an inherent property of the model and not due to specific parameter choices.
- Increasing the rate of adult classifier addition (to simulate innovation) did not significantly alter the generalization trend, further supporting the role of learning.
- Simulations where classifiers were initialized with many features or dropped randomly led to rapid system collapse, which is inconsistent with observed linguistic data, highlighting the importance of the specific learning mechanism and parameterization.
Limitations and Future Work
- Limitations:
- The corpus analysis was limited by the availability of Cantonese CDS data, preventing a direct comparison of child acquisition patterns.
- The simulation uses an abstract representation of semantic features, which doesn't fully capture the nuances of real-world word coining or the creation of new classifiers through grammaticalization.
- The current model doesn't fully incorporate sociolinguistic factors or adult-driven change, which may play a role in classifier death or the adoption of specific classifiers for emphasis.
- Future Work:
- Investigate whether young children use classifiers to disambiguate more often than adults, especially in Mandarin vs. Cantonese.
- Explore the role of discourse factors and adult choices in the replacement of classifiers (e.g., the shift from méi to gè in Chinese).
- Extend the model to incorporate more concrete semantic representations and algorithms for word coining.
- Incorporate sociolinguistic network topology and model both acquisition-driven and sociolinguistic change simultaneously.
Summary updated